WO2020119156A1

WO2020119156A1 - Casting mold breakout prediction method based on feature vectors and hierarchical clustering

Info

Publication number: WO2020119156A1
Application number: PCT/CN2019/100130
Authority: WO
Inventors: 王旭东; 段海洋; 姚曼
Original assignee: 大连理工大学
Priority date: 2018-12-11
Filing date: 2019-08-12
Publication date: 2020-06-18
Also published as: US20210048402A1; CN109396375B; US11105758B2; CN109396375A

Abstract

A casting mold breakout prediction method based on feature vectors and hierarchical clustering. According to the present prediction method, temperature feature vectors of sticking breakout, past data under normal working conditions, and online real-time measured data are extracted to establish a sample set of feature vectors; the sample set is normalized and subject to hierarchical clustering; and then, whether the feature vectors extracted online falls within the breakout cluster is checked and determined, so as to identify and predict casting mold breakout. The prediction method avoids the tedious debugging and modification of parameters such as an alarm threshold, overcomes the artificial dependence of previous breakout prediction methods, and has good robustness and migration. By means of temperature feature extraction, not only can the temperature mode of sticking breakout be accurately recognized, false alarms can be avoided and the number of false alarms can be significantly reduced, the amount of data calculation and calculation time can be greatly reduced, thereby ensuring the instantaneity of online prediction.

Description

A prediction method of mold steel leakage based on feature vector and hierarchical clustering

Technical field

The invention belongs to the technical field of iron and steel metallurgy continuous casting detection, and relates to a method for predicting steel leakage based on feature vectors and hierarchical clustering.

Background technique

Bonding leakage steel refers to the rupture of the thin primary shell near the meniscus during the continuous casting process. After the molten steel oozes out, it will contact with the copper plate of the mold to form a bond. As the mold vibrates and the billet moves down, the bond repeatedly tears- Healed and continued to move down. When it moved out of the crystallizer outlet, it lost the supporting constraints of the copper plate, and the molten steel overflowed, causing steel leakage. Steel leakage not only endangers the safety of on-site operators, but also severely damages continuous casting equipment. At the same time, continuous casting production will be forced to be interrupted, and equipment maintenance and production costs will increase significantly. Therefore, online monitoring and forecasting of steel leakage is always the top priority of continuous casting process control, which is of great significance to ensure the smooth progress of production.

At present, the existing method for predicting steel leakage mainly consists of embedding and installing a thermocouple on the copper plate of the crystallizer, and monitoring and judging whether there is adhesion between the slab and the copper plate according to the temperature change of the thermocouple. Practice has shown that the rate of change of thermocouple temperature with time and its amplitude when bonding occurs are significantly different from normal operating conditions. Therefore, the common characteristics of thermocouple temperature and its rate and amplitude can be extracted and summarized, combined with logic Judgment or neural network method to distinguish and identify typical temperature characteristics of bonding, and online prediction of mold leakage.

Judging from the actual application of the existing steel leakage prediction technology, the steel leakage prediction method based on logical judgment has a strong dependence on continuous casting equipment, process parameters and physical property parameters. When the process adjustment and drawing speed increase, the threshold value changes greatly , Leading to a significant increase in the false positive rate and false negative rate; the neural network method has higher requirements for learning and training samples. When the sample is incomplete or invalid, it will seriously affect its prediction effect, and the model's ability to migrate is low. In practice, in order to avoid missed reports, a certain surplus will be reserved when designing the forecast algorithm or setting the alarm threshold. In the case of process adjustment and equipment maintenance and replacement, the forecast threshold needs to be frequently debugged and corrected manually. If the setting is improper or adjusted If it is not timely, it is difficult to accurately identify and eliminate false alarms, and even lead to omissions. It is a common problem faced by the current steel leakage forecast system.

Patent CN106980729A discloses a method for predicting continuous casting steel leakage based on a hybrid model. This method collects and stores real-time data of all thermocouple temperatures on site, and determines whether the time series of each thermocouple temperature change is consistent with that of bonded steel leakage Temperature change waveform, save the judgment result as Y(i,j,t), if Y(i,j,t) is within the set threshold value, it is marked as thermocouple abnormal, and the current thermocouple line and the previous line are counted For the number of abnormal thermocouples, compare the total number of abnormal thermocouples output with the thresholds for the number of thermocouples for sticking warnings and sticking alarms. This method focuses on the temperature change of a single thermocouple and the number of superimposed thermocouples. The important "time lag" for bonded leaking steel is not involved, making it more accurate to capture the "time lag" and "temperature inversion" characteristics of leaking steel. Difficulties have affected the prediction accuracy of this method.

Patent CN102554171A discloses a continuous casting steel leakage prediction method, which introduces an adaptive algorithm into the BP neural network to achieve automatic optimization of the network structure, and proposes a frictional steel leakage prediction model based on logical judgment and neural network. Combining the accuracy of temperature monitoring with the sensitivity of friction monitoring, a prediction mechanism based on temperature monitoring and supplemented by friction monitoring has been established. The method is provided with four alarm points for temperature monitoring, single-couple, group-couple, friction logic judgment, and sequential network, and they have three levels of yellow, orange, and red. Each alarm point needs to set a corresponding threshold, and there are many forecast parameters. This makes the prediction method of steel leakage prediction poor in universality and mobility.

Patent CN108580827A discloses a method for predicting steel mold leakage based on condensed hierarchical clustering. By establishing a sample library of bonded steel leakage and normal working conditions, an equal amount of samples are randomly selected from each of the sample set of bonded steel leakage and normal working conditions. , And online measured temperature samples constitute a random sample set, implement hierarchical clustering on the random sample set, and then check whether the online measured temperature samples belong to the bonded steel leakage clusters, in order to identify and predict steel leakage. The invention gets rid of the dependence on the artificially defined parameters in the forecasting process, and only uses the characteristics of the bonding leaked steel and the normal operating temperature to determine whether the online measured temperature sample includes the leaked steel. However, the limitation of this method is: 1) randomly select some samples in all steel leakage sample sets for clustering, and cannot take into account the common characteristics of all samples and the individual characteristics of a single sample, that is, the sample characteristics are not fully covered after clustering; 2) temperature The data is directly used for clustering after preprocessing, and the amount of data and calculation are large, which affects the real-time nature of online forecasting. In response to the above problems, on the one hand, the present invention reduces the dimensionality and calculation amount of data through temperature feature extraction. On the other hand, it uses all samples including steel leakage and normal working conditions for hierarchical clustering to take into account the common characteristics and personality characteristics of the samples. , On the premise of ensuring the accuracy of the forecast, improve the online computing speed and real-time.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings of the existing forecasting method in the steel leakage feature extraction, and propose a crystallizer steel leakage forecasting method based on feature vectors and hierarchical clustering. By extracting all the bonded steel leakage samples, normal working conditions Historical data and temperature feature vectors of online measured data establish a feature vector library; normalize and hierarchical cluster the feature vector library; thereafter check and judge whether the feature vectors extracted online belong to the steel leakage cluster, and then identify and Predicting mold leakage.

To achieve the above objectives, the technical solution of the present invention is as follows:

A method for predicting steel mold leakage based on feature vector and hierarchical clustering includes the following steps:

The first step is to extract the characteristic vector

(1) Obtain the historical temperature data of the bonded steel leakage: mark the highest moment of the temperature of the first row of galvanic couples where the bonding position is located, and select the temperature data of M+N seconds before M seconds and N-1 seconds after;

(2) Extract and construct the temperature characteristic vectors of the first and second rows of galvanic couples: each column of galvanic couples is used as a unit to extract the characteristics of the temperature change of the same column of galvanic couples along the casting direction when the steel leakage is bonded, including:

1st_Rising_Amplitude: the temperature rise amplitude of the first row;

1st_Rising_V_Max: the maximum value of the temperature rise rate of the first row;

1st_Falling_V_Ave: mean value of temperature drop rate of the first row;

2nd_Rising_V_Max: the maximum value of the temperature rise rate of the second row;

1st_2nd_Time_Lag: temperature rise time lag, that is, the time interval between the second row and the first row of thermocouple temperature starts to rise;

Construct the feature vector from this:

s = [1st_Rising_Amplitude, 1st_Rising_V_Max, 1st_Falling_V_Ave, 2nd_Rising_V_Max, 1st_2nd_Time_Lag]

The second step is to extract the feature vector of normal working conditions

(1) Obtain historical temperature data of normal working conditions: arbitrarily intercept continuous M+N seconds of temperature data; extract and construct temperature characteristic vectors of the first and second rows of galvanic couples: take each column of galvanic units as the unit to extract normal working Under the condition of the same column galvanic temperature change characteristics along the casting direction, the specific method is the same as the first step (2);

The third step is to extract online real-time temperature features

(1) Real-time collection and acquisition of the inner arc wide face, outer arc wide face, left narrow face, right narrow face copper plate rows and columns of the thermocouple at the current time and before M+N-1 seconds, a total of M+N seconds Temperature data

(2) Extract and construct the temperature characteristic vectors of the first and second rows of galvanic couples: each column of galvanic couples is used as a unit to extract the online real-time temperature characteristics of the same column of galvanic couples along the casting direction. The specific method is the same as the first step (2);

The fourth step is to establish a feature vector library

(1) Establish a feature vector sample library D based on the bonded leakage steel extracted in the first, second and third steps, normal working conditions and online measured temperature characteristics;

(2) The feature vector sample library D is normalized to obtain the feature vector set S, and the normalized online measured temperature feature is recorded as s _new . The specific normalization method is as follows:

Where x _ij is the value of the j-th feature of the i-th feature vector in the feature vector set S, x _jmax and x _jmin are the maximum and minimum values of the j-th feature of all feature vectors, respectively, and |S| represents the feature vector set The total number of vectors in S.

The fifth step, feature vector hierarchical clustering

(1) Implement hierarchical clustering on the feature vector set S obtained in the fourth step, the specific process includes:

1.1) Consider each vector s in the feature vector set S as an initial cluster C _i ={s _i }, and establish a cluster set C={C ₁ ,C ₂ ,...,C _k }; where s _i represents the i-th vector in S, C _i represents the i-th cluster, i=1, 2, ..., k, k represents the total number of vectors in the feature vector set S;

1.2) Calculate and determine the distance between any two clusters C _p and C _q in cluster set C:

d(C _p ,C _q )=min(dist(C _pi ,C _qj ))

Where C _pi is the i-th feature vector in cluster C _p , C _qj is the j-th feature vector in cluster C _q , and dist(C _pi , C _qj ) represents the Euclidean distance of feature vectors C _pi and C _qj ; class cluster computing C _p, C _q from two arbitrary vectors, as taken from the minimum value min clusters by a distance of C _p and C _q.

1.3) Mark the two clusters C _m and C _n with the smallest distance after the calculation in step 1.2), merge C _m and C _n into a new cluster C _{m,n} and add it to the set C, while deleting the original After adding and deleting clusters C _m and C _n , the total number of clusters in set C is reduced by 1;

1.4) Loop execution steps 1.2) to 1.3), when there are only two clusters left in cluster set C, the loop ends and the clustering process is completed;

(2) Check whether the clustering result meets the following judgment conditions, namely:

More than 90% of all bonded steel leakage feature vectors belong to the same cluster, and the ratio of feature vectors under normal operating conditions in this cluster is less than 20%;

If this condition is met, the cluster is classified as a steel _breakout cluster C _breakout , and the other cluster is recorded as a normal operating cluster C _normal ; otherwise, steps (1) and (2) are re-executed until the steel leakage and normal The clustering result of feature vector set composed of operating conditions and measured temperature meets the above judgment conditions;

The sixth step, steel leakage identification and alarm

Determine whether the new feature vector s _new belongs to the cluster C _breakout , if it is, a steel leakage alarm is issued; otherwise, continue to perform the third, fourth, fifth, and sixth steps.

The above method for predicting steel leakage is suitable for the identification of steel leakage in continuous casting slabs such as slabs, square billets, round billets, and shaped billets.

The beneficial effects of the present invention are: A proposed method for predicting steel leakage based on feature vectors and hierarchical clustering avoids the cumbersome debugging and modification of parameters such as alarm thresholds, and overcomes the artificial dependence of previous steel leakage prediction methods It has good robustness and mobility; through temperature feature extraction, it can not only accurately identify the temperature pattern of bonded steel leakage, avoid false alarms and significantly reduce the number of false alarms, but also greatly reduce the amount of data calculation and operation time. Ensure the real-time nature of online forecasts.

BRIEF DESCRIPTION

Figure 1 is a schematic diagram of the distribution of four crystallizer copper plates and thermocouples;

Figure 2 is a schematic diagram of the feature vector extraction of the temperature and the rate of change when the steel leakage is bonded;

Figure 3 is a schematic diagram of feature vector extraction of temperature and its rate of change under normal operating conditions;

Figure 4 is a flowchart of hierarchical clustering of feature vector set and steel leakage recognition and alarm;

Figure 5 is the online measured temperature 1;

Fig. 6 is the results of hierarchical clustering and steel leakage prediction including online measured temperature feature vector 1;

Figure 7 is the online measured temperature 2;

Figure 8 is the result of hierarchical clustering and steel leakage prediction including online measured temperature feature vector 2.

detailed description

The present invention will be further described below with specific embodiments and in conjunction with the accompanying drawings.

The invention is mainly composed of six parts: extracting the characteristic vector of the steel leakage, extracting the characteristic vector of the normal working condition, extracting the online real-time temperature characteristic vector, establishing the characteristic vector library, the hierarchical clustering of the characteristic vectors, and identifying and alarming the steel leakage.

Figure 1 is a schematic diagram of the distribution of four crystallizer copper plates and thermocouples. The slab continuous casting mold is composed of four copper plates, including an outer arc wide surface copper plate, a left side narrow surface copper plate, an inner arc wide surface copper plate, and a right side narrow surface copper plate. The length L is 900 mm, respectively. Two rows of measuring points are arranged on the horizontal cross-section of L1 210mm and L2 325mm, 19 rows of thermocouples are arranged in each row on the outer arc wide copper plate and the inner arc wide copper plate, the distance between the two galvanic couples L3 is 150mm, two There are 38 thermocouples on each of the wide-surface copper plates; a row of thermocouples are arranged on the center line on the left narrow-surface copper plate and a right-side narrow-surface copper plate, and two thermocouples are arranged on each of the two narrow-surface copper plates. The total number of galvanic couples arranged by four copper plates is 80, and the distance from each galvanic couple to the hot surface of the copper plate of the crystallizer is equal.

Step 1: Extract the characteristic vector of cemented steel leakage

(1) Obtain the historical temperature data of the bonding leakage steel: mark the moment of the highest temperature of the first row of galvanic couples at the bonding position, and select the temperature data of the first 15 seconds and the last 9 seconds for a total of 25 seconds;

For the historical temperature of bonded steel leakage, 30 temperature samples were selected.

(2) Extract and construct temperature characteristic vectors of the first and second rows of galvanic couples

Fig. 2 is a schematic diagram of feature vector extraction of the temperature of the bonding leak steel and its temperature change rate. It can be seen from FIG. 2 that the feature vectors are extracted in units of each column of galvanic couples, and the characteristics of the temperature changes of the same column of galvanic couples along the casting direction when extracting the bonded leaked steel are extracted. The specific features and extraction methods are as follows:

2.1) Calculate the change rate of temperature data at the same measuring point within 5 seconds, namely:

In the formula, r ∈ [1,2] represents the first and second rows of thermocouples respectively, T _(r)i represents the temperature value of the r-th row thermocouple at time i, and v _(r)i represents the r- _th row thermoelectric The value of the temperature change rate at time i.

2.2) Determine the starting temperature and corresponding time when the temperature rises:

First, obtain the maximum value T _max including the current time and the temperature within 25 seconds before it and its corresponding time t _max , namely:

T _max = max(T _i ), i=1, 2, ..., 25,

Then, traverse forward from T _max to obtain the minimum value of the previous temperature T _min and the time t _min , that is:

T _min =min(T _i ),i=1, 2,..., t _max

Then take T _min as the initial temperature when the temperature rises, and the corresponding time is recorded as t _min .

2.3) Extract corresponding features to construct feature vectors:

1st_Rising_Amplitude: the first row of temperature rise amplitude, that is, the temperature at which the first row of temperature starts to rise and reaches the maximum, respectively, and calculate the difference between the two temperatures to obtain the temperature rise amplitude in ℃:

1st_Rising_Amplitude=T _(1)max -T _(1)min

1st_Rising_V_Max: the maximum value of the first row of temperature rise rate, that is, the maximum value of the temperature change rate of the first row of thermocouples obtained in the extraction step 2.1), in ℃/s:

1st_Rising_V_Max=max(v ₍₁₎ )

1st_Falling_V_Ave: the average value of the temperature drop rate of the first row, that is, the average value of the temperature rate of the temperature change rate of the first row thermocouple less than 0 obtained in the calculation step 2.1), the unit is ℃/s;

2nd_Rising_V_Max: the maximum value of the temperature rise rate of the second row, that is, the maximum value of the temperature change rate of the second row thermocouple obtained in the extraction step 2.1), in ℃/s;

2nd_Rising_V_Max=max(v ₍₂₎ )

1st_2nd_Time_Lag: Temperature rise time lag, that is, mark the time when the temperature of the second row and the first row start to rise, and take the time interval of the two as the temperature rise time lag, the unit is s:

1st_2nd_Time_Lag=t _(2)min -t _(1)min

Construct the feature vector from this:

Step 2: Extract feature vectors under normal operating conditions

(1) Obtain historical temperature data under normal working conditions: arbitrary interception of continuous 25-second temperature data;

Figure 3 is a schematic diagram of feature vector extraction of temperature and its temperature change rate under normal operating conditions. It can be seen from FIG. 3 that the feature vectors are extracted in units of each column of galvanic couples to extract the characteristics of the temperature changes of the same column of galvanic couples along the casting direction under normal operating conditions. The features and their extraction methods are the same as in step one (2).

For normal operating temperature, 30 temperature samples were selected.

Figure 4 is a flowchart of feature vector set hierarchical clustering and steel leakage recognition and determination. It can be seen from the figure that the hierarchical clustering of feature vector sets and the identification and determination of steel leakage mainly include the following steps:

Step 3: Extract online real-time temperature feature vector

(1) Collect and acquire real-time temperature data of the inner arc wide face, outer arc wide face, left narrow face, and right narrow face copper plate of each row and column of the thermocouple at the current time and 24 seconds before, 24 seconds in total;

(2) Extract and construct the temperature characteristic vectors of the first and second rows of galvanic couples:

The feature vectors are extracted in units of each column of galvanic couples, and the characteristics of the temperature changes of the same column of galvanic couples along the casting direction during online measurement are extracted. The features and their extraction methods are the same as in step one (2).

Step four: establish a feature vector library

(1) Create a feature vector sample library D, a total of 61 cases, based on the bonded leakage steel, normal working conditions and online measured temperature features extracted in

steps

1, 2, and 3.

(2) Normalize the feature vector sample library D to obtain the feature vector set S, and record the normalized online measured temperature feature vector as s _new . The specific normalization method is as follows:

Where x _ij is the value of the j-th feature of the i-th feature vector in the feature vector set S, and x _jmax and x _jmin are the maximum and minimum values of the j-th feature of all 61 feature vectors, respectively.

Step 5: Hierarchical clustering of feature vectors

(1) Implement hierarchical clustering on the feature vector set S obtained in step 4;

1.1) Treat each vector s in the feature vector set S as an initial cluster C _i ={s _i }, establish a cluster set C={C ₁ ,C ₂ ,...,C ₈₁ }; where s _i represents the i-th vector in S, C _i represents the i-th cluster, i=1, 2, ..., 61.

d(C _p ,C _q )=min(dist(C _pi ,C _qj ))

Step 6. Leakage identification and alarm

Determine whether the new feature vector s _new belongs to the cluster C _breakout , if it is, a steel leakage alarm is issued; otherwise, continue to perform steps three, four, five, and six.

Fig. 5 shows the temperature diagrams of the first and second rows of galvanic couples measured online at temperature 1. The vertical line on the right side of the figure shows the current time when online detection is performed, and the 24 seconds before this time, a total of 25 seconds of temperature data. The feature vector obtained after feature extraction for online temperature 1 is:

s _new = [0.4,0.18,-1.18,1.72,0],

Fig. 6 is a hierarchical clustering of the feature vector set containing online measured temperature feature vectors, that is, a feature vector set containing s _new and a prediction result of steel leakage. It can be seen from the figure that after normalization and hierarchical clustering, the feature vector set is clustered into two clusters: the cluster on the left contains all the bonded leaking steel feature vectors labeled 1 and 5 labels 2 Normal working condition samples, the ratio of the characteristic vector of bonded steel leakage in this type of cluster is greater than 90% of the total number of samples of bonded steel leakage, the ratio of the characteristic vector of normal working conditions is less than 20% of the total number of samples of the normal working condition, that is, 6 The feature vector meets the criteria for cluster classification, and the hierarchical clustering of the feature vector set is successful, so this cluster is recorded as a steel leakage cluster C _breakout ; another cluster with a large number of 2 is recorded as a normal working cluster C _normal . It can also be seen from Fig. 6 that the online measured temperature 1 feature vector s _new obtained by feature extraction, that is, the sample labeled "N", belongs to the normal working condition cluster C _normal after clustering, and does not belong to the bonding leak Steel cluster C _breakout , so it is judged as normal working condition, then continue to update the temperature sequence, perform steps three, four, five, six.

Fig. 7 shows the temperature diagrams of the first and second row galvanic couples of the temperature 2 measured online. The vertical line on the right side of the figure shows the current time when online detection is performed, and the 24 seconds before this time, a total of 25 seconds of temperature data. The feature vector obtained after feature extraction for online temperature 2 is:

s _new = [7.4,1.06,-0.29,1.36,12],

Fig. 8 is a graph containing hierarchical clustering and steel leakage prediction results of the online measured temperature feature vector 2, that is, the feature vector set including s _new . It can be seen from the figure that after normalization and hierarchical clustering, the feature vector set is clustered into two clusters: the cluster on the left contains all the bonded leaked steel samples labeled 1 and 5 labeled 2 In normal working condition samples, the percentage of the characteristic vector of bonded steel leakage in this cluster is greater than 90% of the total number of samples of the steel leakage, and the ratio of the characteristic vector of normal operating conditions is less than 20% of the total number of samples in the normal working condition, that is, 6 features The vector meets the criteria for cluster determination, and the hierarchical clustering of the feature vector set is successful, so this cluster is recorded as a steel leakage cluster C _breakout ; then another cluster with a large number of 2 is recorded as a normal operating cluster C _normal . It can also be seen from Figure 8 that the online measured temperature 2 feature vector s _new obtained by feature extraction, that is, the sample labeled "B", belongs to the bonded steel leakage cluster C _breakout after clustering, so it is determined to be steel leakage , A steel leakage alarm is issued.

The above-mentioned examples only express the embodiments of the present invention, but they should not be construed as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, without departing from the concept of the present invention, Several modifications and improvements can also be made, which all fall within the protection scope of the present invention.

Claims

A method for predicting mold steel leakage based on feature vectors and hierarchical clustering, characterized in that the prediction method extracts temperature feature vectors of bonded steel leakage, historical data of normal working conditions and online measured data, respectively, to establish a sample set of feature vectors; Normalize the sample set and perform hierarchical clustering; thereafter check and judge whether the feature vectors extracted online belong to the steel leakage cluster, and then identify and predict the mold leakage, including the following steps:

The first step is to extract the characteristic vector

(1) Obtain the historical temperature data of the bonded steel leakage: mark the highest moment of the temperature of the first row of galvanic couples where the bonding position is located, and select the temperature data of M+N seconds before M seconds and N-1 seconds after;

(2) Extract and construct temperature characteristic vectors of the first and second rows of galvanic couples;

The second step is to extract the feature vector of normal working conditions

(1) Obtain historical temperature data under normal working conditions: arbitrary interception of continuous M+N seconds of temperature data;

(2) Extract and construct temperature characteristic vectors of the first and second rows of galvanic couples;

The third step is to extract online real-time temperature features

(1) Real-time collection and acquisition of the inner arc wide face, outer arc wide face, left narrow face, right narrow face copper plate rows and columns of the thermocouple at the current time and before M+N-1 seconds, a total of M+N seconds Temperature data

(2) Extract and construct temperature characteristic vectors of the first and second rows of galvanic couples;

The fourth step is to establish a feature vector library

(1) Establish a feature vector sample library D based on the bonded leakage steel extracted in the first, second and third steps, normal working conditions and online measured temperature characteristics;

(2) Normalize the feature vector sample library D to obtain the feature vector set S, and record the normalized online measured temperature feature as s new ;

The normalization method of the feature vector is as follows:

Where x ij is the value of the j-th feature of the i-th feature vector in the feature vector set S, x jmax and x jmin are the maximum and minimum values of the j-th feature of all feature vectors, respectively, and |S| represents the feature vector set The total number of vectors in S;

The fifth step, feature vector hierarchical clustering

(1) Implement hierarchical clustering on the feature vector set S obtained in the fourth step, the specific process includes:

1.1) Consider each vector s in the feature vector set S as an initial cluster C i ={s i }, and establish a cluster set C={C 1 ,C 2 ,...,C k }; where s i represents the i-th vector in S, C i represents the i-th cluster, i=1, 2, ..., k, k represents the total number of vectors in the feature vector set S;

1.2) Calculate and determine the distance between any two clusters C p and C q in cluster set C:

d(C p ,C q )=min(dist(C pi ,C qj ))

Where C pi is the i-th feature vector in cluster C p , C qj is the j-th feature vector in cluster C q , and dist(C pi , C qj ) represents the Euclidean distance of feature vectors C pi and C qj ; class cluster computing C p, C q from two arbitrary vectors, as taken from the minimum value min clusters by a distance of C p and C q;

1.3) Mark the two clusters C m and C n with the smallest distance after the calculation in step 3.2), merge C m and C n into a new cluster C {m,n} and add it to the set C, while deleting the original After adding and deleting clusters C m and C n , the total number of clusters in set C is reduced by 1;

1.4) Loop execution steps 3.2) to 3.3), when there are only two clusters left in cluster set C, end the loop and complete the clustering process;

(2) Check whether the clustering result meets the following judgment conditions, namely:

More than 90% of all bonded steel leakage feature vectors belong to the same cluster, and the ratio of feature vectors under normal operating conditions in this cluster is less than 20%;

If this condition is met, the cluster is classified as a steel breakout cluster C breakout , and the other cluster is recorded as a normal operating cluster C normal ; otherwise, steps (1) and (2) are re-executed until the steel leakage and normal The clustering result of feature vector set composed of operating conditions and measured temperature meets the above judgment conditions;

The sixth step, steel leakage identification and alarm

Determine whether the new feature vector s new belongs to the cluster C breakout , and if it is, issue a steel leakage alarm; otherwise, continue to perform the third, fourth, fifth, and sixth steps;

The first step (2), the second step (2) and the third step (2) involve the same temperature feature extraction method, taking each column as a unit to extract the temperature along the same column under different working conditions Variation characteristics of casting direction, including:

1st_Rising_Amplitude: the temperature rise amplitude of the first row;

1st_Rising_V_Max: the maximum value of the temperature rise rate of the first row;

1st_Falling_V_Ave: mean value of temperature drop rate of the first row;

2nd_Rising_V_Max: the maximum value of the temperature rise rate of the second row;

1st_2nd_Time_Lag: temperature rise time lag, that is, the time interval between the second row and the first row of thermocouple temperature starts to rise;

Construct the feature vector from this:

s=[1st_Rising_Amplitude, 1st_Rising_V_Max, 1st_Falling_V_Ave, 2nd_Rising_V_Max, 1st_2nd_Time_Lag].
The method for predicting steel leakage based on feature vectors and hierarchical clustering according to claim 1, characterized in that the method for predicting steel leakage is suitable for continuous casting of slabs, square billets, round billets and shaped blanks Online leak prediction of the process.